text stringlengths 0 59.1k |
|---|
- LLM: Sends a prompt to a selected model and returns structured or free form output. |
- Memory: Reads and writes scoped state for later steps. |
- Output: Produces the terminal response or emits an event. |
Beyond the node palette, the builder now emphasizes versioned workflows. Each publish creates a snapshot that can be pinned in ChatKit or exported as SDK code. Preview runs show trace by trace data, and the Evaluate tab lets teams run trace graders without leaving the canvas. Templates across support, research, and int... |
In practice, this shape works well for support bots, data lookups, and lightweight automations, the kind of flows people prefer to see end to end on one screen. |
## AgentKit Adds More Building Blocks |
AgentKit positions the builder as one pillar in a package that covers orchestration, observability, and deployment. Three parts stand out when piecing together flows: |
- Connector Registry stabilizes the security posture. It is a single admin surface where ChatGPT workspaces and API organizations map to data sources like Google Drive, SharePoint, Microsoft Teams, Dropbox, plus any MCP connector the admin approves. No separate spreadsheet of systems is required. |
- ChatKit handles the front end. It takes a published workflow ID, manages threads, streams responses, and shows the agent thinking indicator without rebuilding the chrome yet again. Version swaps happen server side, which keeps the review loop fast. |
- Guardrails provides the safety layer. Whether mounted inside the builder or alongside a VoltAgent runtime, it offers jailbreak checks, PII masking, and policy hooks that fit into the node graph. |
Taken together, AgentKit handles proof of concept graphing, the UI surface, and the compliance checklist larger organizations need. VoltAgent steps in when teams want handwritten orchestration, custom memory strategies, or the option to mix model providers beyond what the hosted stack offers. |
## Integration Points |
On integrations, Agent Builder still connects to the OpenAI stack. MCP handles typed tool calls, ChatKit provides UI components, the editor taps directly into OpenAI APIs, and there is a one step deploy to the hosted runtime. In practice that means less glue code between a sketch and something teammates can click. |
Publishing yields a versioned ID. Teams can keep an early version pinned for production while experimenting with a branch in preview, then roll forward only after the built in graders pass. If there is a need to own the runtime, the builder will export TypeScript that mirrors the graph so the workflow can move into Vol... |
## Guardrails and Constraints |
On the safety side, the runtime supports constraints to contain behavior: |
- Deny lists for specific data sources or HTTP targets. |
- Approval gates that pause execution until a user confirms an action. |
- Response checks that reject outputs that violate a content policy. |
I like that constraints sit in the graph like regular nodes, which makes them easy to reason about during testing and easy to diff during reviews. |
## Evaluate and Improve |
OpenAI’s latest update adds evaluation-first tooling into the same workflow. Teams can attach a curated set of traces, run the automated graders, and tweak prompts without leaving the canvas. For trickier problems it is possible to lean on reinforcement fine tuning. RFT is generally available on o4-mini and in private ... |
## So |
If you’re starting cold, the templates help. Today’s set covers support, content generation, research, and internal automations. They’re regular flows you can edit, extend, or throw away, which keeps them useful as scaffolds rather than prescriptions. |
VoltAgent is an open-source TypeScript framework for orchestrating AI agents. It gives you control over workflows, sub-agents, memory adapters, and observability. VoltAgent is built for production with support for retries, tracing, error handling, and type safety. You can integrate it with any front end or use it behin... |
- GitHub repo: [VoltAgent/voltagent](https://github.com/VoltAgent/voltagent) |
- Examples & templates: [VoltAgent Examples](https://voltagent.dev/examples/) |
Also part of the VoltOps ecosystem: VoltOps handles deployment, monitoring, and operations around VoltAgent agents. |
As AgentKit matures, VoltAgent stays the option when teams need richer customization, multi provider routing, or on premise deployments. Draft flows in the builder, export them, and evolve the logic in VoltAgent while stakeholders continue collaborating on the visual canvas. |
<|endoftext|> |
# source: VoltAgent__voltagent/website/blog/2025-04-25-what-is-an-mcp-server/index.md type: docs |
--- |
title: What's MCP and Why Should I Care? |
description: Learn about the Model Context Protocol (MCP) and how it helps AI agents like VoltAgent interact with external tools, with a practical example. |
slug: what-is-mcp |
tags: [mcp] |
authors: omeraplak |
image: https://cdn.voltagent.dev/2025-04-25-what-is-an-mcp-server/social.png |
--- |
import Tabs from '@theme/Tabs'; |
import TabItem from '@theme/TabItem'; |
import GitHubExampleLink from '@site/src/components/blog-widgets/GitHubExampleLink'; |
## Introduction |
Ever built an AI agent and wondered how it actually _does_ stuff in the real world? Like, how does it browse the web, read a file from your computer, or talk to a database? |
It turns out a key piece of the puzzle is something called the **Model Context Protocol**, or **MCP**. It sounds technical, but the core idea is pretty neat and solves a big problem. |
In this post, I'll walk you through: |
- [What is MCP](#what-is-mcp) |
- [Why Should I Care About MCP?](#why-should-i-care-about-mcp) |
- [Finding MCP Servers](#finding-mcp-servers) |
- [Introducing VoltAgent](#introducing-voltagent) |
- [VoltAgent and MCP](#voltagent-and-mcp) |
- [Let's Build an Example with VoltAgent!](#lets-build-an-example-with-voltagent) |
- [Setting Up the Project](#setting-up-the-project) |
- [Implementing the Agent and MCP Configuration](#implementing-the-agent-and-mcp-configuration) |
- [Running the Agent](#running-the-agent) |
- [Testing in the Console](#testing-in-the-console) |
## What is MCP? |
I like to think of an AI agent as a very smart brain in a jar. It's brilliant at understanding and generating language, but it's stuck inside its digital container. It doesn't have built-in hands or eyes to interact directly with the outside world files, websites, databases, APIs, etc. To do useful tasks, it needs **to... |
:::info Think of it like this: |
Imagine you have a bunch of different power tools: a drill, a saw, a sander. If each one needed a completely different type of power cord and plug, it would be a nightmare. |
MCP acts like a **universal adapter** or a **standard plug socket** for AI tools. |
::: |
Instead of every tool (like a file reader or a web browser) needing a unique, custom connection to the AI agent, MCP provides a **standard way** for agents and tools (which we call "MCP servers") to communicate. |
This means an agent I build with a framework that understands MCP can potentially connect to _any_ service or tool that also speaks this standard MCP language. This could be a tool for: |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.